ARMOR: Pareto-Aware Dynamic Routing for Energy-Efficient Heterogeneous Multimodal Perception in Autonomous Driving

Abstract

Modern autonomous driving systems are constrained by the tension between high-fidelity perception and stringent onboard energy conservation. We propose adaptive routing of multimodal onboard reasoning (ARMOR), an end-to-end heterogeneous mixture-of-experts framework that dynamically routes inputs between computationally efficient pure-vision experts and robust radar-vision fusion experts. A lightweight context-aware gating network, the multimodal adaptive routing sentinel (MARS), integrates temporal context and environmental priors to select appropriate experts. ARMOR was optimized using a unified cost-conditioned multiobjective loss, enabling a single model to learn routing policies across an empirical cost–performance trade-off frontier and adapt to different operational preferences. To improve reliability in high-uncertainty scenarios, an evidential-learning-based sentry mechanism was used to trigger fallback to a high-fidelity fusion expert. Experiments on the simulation-generated CARLA-SUMO multimodal object detection dataset (CSODD) showed that ARMOR-safety retains 98.1% of the mean average precision of the strongest fusion baseline while reducing average floating-point operations per second (FLOPs), latency, and the profiled GPU-side inference-energy proxy by 66.3%, 64.4%, and 65.8%, respectively. In adverse weather, it improved mean average precision by 58.3% over high-performance pure-vision models and reduced the critical-object miss rate by more than 50%. These results provide controlled framework-level evidence, but do not establish cross-dataset generalization or real-road deployment performance.